Case study · 2025
Soccer Match Highlights Event Extraction
Built a multimodal AI pipeline that detects key match events and generates structured highlight outputs from full-length football videos.
engineeringresearch
Overview
This capstone project focused on turning long-form sports footage into structured event data and clipped highlights. The system was designed to reduce manual review time and create reusable outputs for downstream applications.
Pipeline
- OCR extracted scoreboard and visual match context from video frames.
- Audio processing identified likely event windows.
- A CRNN architecture with attention was used for audio event classification.
- FFmpeg handled processing and highlight clip generation.
- FastAPI exposed the output pipeline.
Technical Decisions
The project used a multimodal approach because audio or video alone was not reliable enough for practical event extraction. Combining both signals improved confidence and made the output more useful.
Outputs
The system generated structured JSON metadata and clipped highlight segments based on detected event timestamps.